amisima/Qwen-Image-Edit-2511-Custom-Lora-Fast

Qwen-Image-Edit 2511 · Rapid AIO · Custom lora

Edit, relight, re-angle or upscale an image with a style adapter, a fast adapter, and up to three lora of your own.

Please read before uploading a photograph

Upload only photographs of yourself, or of people who have agreed to this. Sexual imagery of a real person, made without that person's consent, breaks Hugging Face's [content policy](https://huggingface.co/content-policy) and is not welcome here. The same applies to anything involving a minor, in any form, without exception.

This Space is tagged Not For All Audiences: adapters loaded into it may produce adult material, and what a visitor loads is their own responsibility under that policy.

What is in here

Bundled adapters, kept private. The catalogue of bundled adapters — their names, their prompts and their trigger words — is not in this repository. It is read at start-up from a private repo named by QIE_CATALOG_REPO; with the variable unset the Space starts with none and the custom slots still work.

Three custom lora slots. A Hugging Face repo, a file inside one, a CivitAI download link, or an uploaded .safetensors. Links are fetched before the GPU is booked and cached, and CivitAI's domains are tried in turn, because one of them will sometimes answer a datacentre address with a Cloudflare challenge while another does not. A gated model needs CIVITAI_TOKEN: the Space downloads on its own machine, so being signed in to CivitAI in a browser does not authenticate it.

Search CivitAI from inside the Space, with size, downloads, base model and trigger words on every row, and one press to drop a result into a slot.

Fast adapters. Step-distilled Lightning adapters that render in 4–8 steps. The Rapid AIO merge is already distilled, so start at None.

A prompt builder that writes the edit instruction around your words, including the clause that holds the subject's face, hair and proportions steady — without it the model quietly turns a person into somebody else.

A live ZeroGPU cost readout, named profiles, and a randomize seed box.

Notes

• —Adapters are converted at load time: kohya naming, the per-head interleaved QKV, the swapped halves of the gated MLP, and alpha folded into the weights.
• —Guidance above 1.0 doubles the passes and the time with it; this merge is tuned for 1.0.